Upload
others
View
7
Download
0
Embed Size (px)
Citation preview
Makara J. Technol. 24/3 (2020), 149159
doi: 10.7454/mst.v24i3.3944
December 2020 | Vol. 24 | No. 3 149
ISO 9001:2015 Risk-based Thinking:
A Framework using Fuzzy-Support Vector Machine
Ralph Sherwin A. Corpuz
Electronics Engineering Technology, Technological University of the Philippines, Manila 1000, Philippines
*e-mail: [email protected]
Abstract
Risk-based thinking (RBT) is one of the distinct new features of the International Organization for Standardization
9001:2015. Interestingly, the standard does not prescribe any tools. Hence, organizations are puzzled as to the extent of
conformance. Some organizations have adopted formal tools. However, these tools seem insufficient in linking the
standard into an evidence-based decision support system. To resolve gaps in RBT implementation, this paper proposes a
framework based on fuzzy inference system (FIS) and support vector machine (SVM) to automate risk analysis and
evaluation, proposal and verification of action plans, and prediction of the feasibility of risks and opportunities according
to text patterns. Modeling results indicate that the framework has no significant difference in terms of accuracy compared
with the conventional method. Both FIS-1 and FIS-2 models, however, are statistically significantly faster at 3.26 and 1.15
s, respectively. Meanwhile, the SVM model, whose text classification features are not evident in the conventional method,
has a 97.16% classification accuracy and 2.6% confusion error during training, and 95% classification accuracy during
testing. Results affirm that FIS and SVM are efficient tools in feasibly conforming with the RBT requirements of the ISO
9001:2015 international standard.
Abstrak
Pemikiran Berbasis Risiko ISO 9001:2015: Suatu kerangka Kerja yang Menggunakan Mesin Vektor Pendukung
yang Kabur. Pemikiran berbasis risiko (Risk-based thinking (RBT)) merupakan salah satu dari fitur-fitur baru yang
berbeda dari Organisasi Internasional untuk Standarisasi 9001:2015 (International Organization for Standardization
9001:2015). Yang menarik, standar tidak menentukan perkakas apapun. Oleh karenanya, berbagai organisasi dibingungkan
dengan tingkat kesesuaian. Sebagian organisasi telah mengadopsi perkakas formal. Namun demikian, perkakas ini
nampaknya tidak mencukupi dalam menghubungkan standar ke dalam suatu sistem pendukung keputusan berbasis
kejadian. Untuk menyelesaikan adanya celah-celah di dalam implementasi RBT, naskah ini mengusulkan suatu kerangka
kerja berdasarkan pada sistem dugaan yang kabur (fuzzy inference system (FIS)) dan mesin vektor pendukung (support
vector machine (SVM)) untuk mengotomatisasi analisis risiko dan evaluasi, usulan dan verifikasi rencara aksi, dan prediksi
kelayakan risiko serta peluang yang sesuai dengan pola-pola naskah. Hasil-hasil pemodelan menunjukkan bahwa kerangka
kerja tersebut tidak memiliki perbedaan yang signifikan dalam hal akurasi dibandingkan dengan metode konvensional.
Namun demikian, baik model FIS-1 maupun FIS-2, secara statistik jauh lebih cepat pada masing-masing 3,26 dan 1,15
detik. Sementara, model SVM, yang fitur-fitur klasifikasi naskahnya bukan kejadian di dalam metode konvensional,
memiliki akurasi klasifikasi 97,16% dan kesalahan yang membingungkan 2,6% selama pelatihan, dan akurasi klasifikasi
95% selama pengujian. Hasil-hasilnya menegaskan bahwa FIS dan SVM merupakan perkakas yang efisien dengan
penyesuaian yang layak dengan persyaratan RBT dari standar internasional ISO 9001:2015.
Keywords: artificial intelligence, fuzzy inference system, ISO 9001:2015, risk-based thinking, support vector machine
1. Introduction
The ISO 9001 Quality Management Systems
Requirements (QMS) International Standard is the most
widely sought standard in the world used for the
attainment of certifications related to quality [1]. The
standard is constantly updated after years of review and
improvement to consistently meet the growing demands
of the global markets for quality products and services.
One of the new distinct features in the ISO 9001:2015
version is the introduction of risk-based thinking (RBT),
which is a conceptual framework that requires an
organization to understand its context through the
determination of internal and external issues and
Corpuz
Makara J. Technol. 1 December 2020 | Vol. 24 | No. 3
150
relevant interested parties; to determine the risks as a
basis for planning activities; to implement QMS
processes; to determine the extent of relevant
documented information; to evaluate and review the
effectiveness of actions to address risks and
opportunities; and to update the risks and opportunities
during planning if necessary [2].
A risk is defined as the effect of uncertainty, which can
either be negative or positive. Mentioned alongside risks
are the possibility of opportunities, which are a set of
circumstances that make it possible to do something
beneficial for an organization. In this context, RBT is
helpful in managing these risks and opportunities to
avoid nonconformities in the future [2]−[4]. To
complement the ISO 9001 requirements, ISO has
published the ISO 31000-Risk Management Principles
and Guidelines, the process map of which is illustrated
in Figure 1 [5] as a guide for organizations in managing
risks.
Interestingly, the standard offers greater flexibility in
implementing the RBT in terms of documented
information, scope of application, and on organizational
roles [2]. However, this flexibility causes confusion
among organizations, particularly with regard to the
extent of the minimum documentation and
implementation requirements [4], [6]. Most service-
oriented multinational organizations implement a formal
risk management process wherein they use traditional
tools aside from the ISO 31000, such as Total
Productive Maintenance, AS/NZS 4360 Risk
Management Standard, and Failure Modes and Effects
Analysis [4], [6]−[7]. Unfortunately, these tools are
incapable of capturing the uncertainties of risk
management process, which are usually expressed in
qualitative data and unrealistically show similar levels
of risk ratings despite different significance levels of
parameters [8].
Figure 1. ISO 31000 Risk Management Process
Artificial intelligence (AI) is an emerging field of
disruptive technologies explored for the design of data-
driven risk management tools. Fuzzy inference system
(FIS) and support vector machines (SVM) are among
the highly sought techniques for such purpose. FIS is a
popular AI technique introduced by Zadeh [9] and is
intended specifically as a measure of uncertainties,
vagueness, and imprecisions. Figure 2 [10] shows a
sample FIS process model, which has five basic
subprocesses. Initially, both input and output data are
modeled into certain membership functions (MF). The
resulting fuzzy sets are then applied with logical
operators whether to intersect (AND) or disjoint (OR)
and then implied to follow a rule-based system, which is
also known as fuzzy rules. Afterward, the resulting
fuzzy sets are aggregated and then finally defuzzified to
yield the desired outputs. FISs are used for
semiquantitative or qualitative types of prediction and
control systems [8], [10]−[11].
SVM is a popular AI technique postulated by Vapnik
[12] based on statistical principles of Huber’s regression
theory and Wolfe’s dual program theory [13]. Figure 3
shows a sample SVM model used for binary
classification [14]. SVM classifies data by finding the
best hyperplane that separates all data points of one
class from the other class. A hyperplane is considered
“best” if it has the largest “margin” between two classes.
The data points that are closest to the “separating
hyperplane” located in the boundary of the margin slab
are called “support vectors.” The “+” and “−” indicate
“type 1” and “type −1” data points, respectively. SVMs
have comparable or higher performance than traditional
learning techniques in terms of generalization ability,
Figure 2. FIS Process Model
Figure 3. SVM Binary Classification Model
ISO 9001:2015 Risk-based Thinking using Fuzzy-Support Vector Machine
Makara J. Technol. 1 December 2020 | Vol. 24 | No. 3
151
robustness, and prediction accuracy. Hence, they are a
popular choice for prediction and classification
applications [15]−[18].
Inevitably, FIS and SVM also have imperfections. FIS
is not efficient in purely quantitative measurements and
in large problems [15]. The same problem is found with
the SVM, which is dependent on its kernel optimization
and is susceptible to slow performance with large
datasets due to memory constraints [19]. In an effort to
resolve these issues, the fusion of FIS and SVM
techniques has opened the feasibility of complementing
their respective strengths and weaknesses to improve
machine learning performance. Extensive studies have
been conducted relative to the use of FIS and SVM,
collectively called fuzzy-SVM, for risk management
systems such as those applied for road safety [15], credit
risk evaluation [18], project risk assessment [19], and
fire risk assessment [20]. These studies, however, are
focused on enhancing optimization techniques, such as
improvement of MF or kernel parameters, and are not
intended as a practical solution to RBT problems. In
fact, there is a dearth of published research papers
related to the use of FIS and SVM for RBT. With that
said, this paper chronicled the design of a framework to
feasibly meet the documentation and implementation
requirements of the ISO 9001:2015 international
standard with respect to RBT. The author focused on
resolving gaps in data entry automation to ensure
efficiency, timeliness, and accuracy of monitoring,
measurement, analysis, and evaluation, and to facilitate
a predictive decision support system for top management
through the use of FIS and SVM techniques.
2. Methods
In this paper, the author utilized RBT data of a state-run
university, the Technological University of the
Philippines (TUP) in Manila, Philippines, as of March
2020. Figure 4 shows the actual RBT process observed
in the university, also known as internal and external
issues assessment procedure. Each office is represented
by a process owner (PO)—who is either a dean,
director, or head—who maintains an online log to
comply with the RBT requirements and updates the
contents at least once a year. However, the certifying
body found that this procedure should be simplified and
should have provisions to monitor the effectiveness of
actions to address risks and opportunities.
As a potential solution to the research gap, the proposed
RBT framework is shown in Figure 5. It is composed of
three complementary modules, namely, the (1) corpus
module, (2) FIS module, and (3) SVM) module. The
corpus module is a collection of textual statements of
issues “I”, management controls “C”, and risks “Ri” and
opportunities “Op.” Table 1 shows a sample corpus
dataset predetermined by a concerned PO.
Figure 4. Sample RBT Process
Figure 5. Proposed RBT Framework
Table 2. Sample RBT Corpus Data
I C
Outdated instructional
tools and equipment
Level 4-Conduct
periodic follow up with
department heads
Ri Op
Inadequate knowledge of
new technologies and
their applications
Trainings and workshops
for teachers
The FIS module was designed using MATLAB with
three distinct objectives: (1) to analyze risks “R” and
opportunities “O”; (2) to determine action plans to
manage risks “RA” and opportunities “OA”; and (3) to
evaluate the effectiveness of action plans to address
risks “RE” and opportunities “OE”. To realize these
objectives, the author utilized equations to standardize
the input data and then modeled two Mamdani-type FIS
models known as FIS-1model, to cater to objectives 1
and 2, and FIS-2 model, to cater to objective 3.
Initially, each PO was required to determine the
equivalent numerical levels of “R” and “O” by rating
the specific parameters by using a five -point scale
criteria. The following formulas were used to quantify
the values of “R,” “O,” and their respective parameters:
𝑅 = 𝑃𝑅𝑆 (1)
where R is the level of risk, and PR is the average
probability and S is the average severity, respectively,
which are further expressed using the following
equations (2 and 3), respectively:
Corpuz
Makara J. Technol. 1 December 2020 | Vol. 24 | No. 3
152
𝑃𝑅 = [𝑃𝑅1+𝑃𝑅2
2] (2)
𝑆 = [𝑆1+𝑆2+𝑆3+𝑆4+𝑆5+𝑆6
6] (3)
With the use of the five-point scale criteria, “PR1”
(previous risk probability) and “PR2” (future risk
probability) were rated as 1 = improbable; 2 = remote;
3 = occasional; 4 = probable; 5 = frequent; and “S1”
(inability to meet customer requirements), “S2”
(potential violation of statutory requirements), “S3”
(potential violation of regulatory requirements), “S4”
(potential violation of organizational policies), and “S5”
(potential impact on organizational reputation) were
rated as 1 = N/A or negligible; 2 = minor; 3 = serious;
4 = critical; 5 = catastrophic; while “S6” (estimated cost
of correction) was rated as 1 = 0 or N/A;
2 = <PhP100,000; 3 = PhP100,001–500,000;
4 = PhP500,001–1,000,000; 5 = >PhP1M.
Afterwards, the PO analyzed the values of “O” by using
the following equation:
𝑂 = 𝑃𝑂𝐵 (4)
where O is the level of opportunity. PO (average
probability) and B (average benefits) are expressed
further in equations (5) and (6), respectively
𝑃𝑂 = [𝑃𝑂1+𝑃𝑂2
2] (5)
𝐵 = [𝐵1+𝐵2+𝐵3+𝐵4+𝐵5+𝐵6
6] (6)
With the use of a five-point scale criteria, “PO1”
(previous opportunity probability) and “PO2” (future
opportunity probability) were rated as 1 = improbable;
2 = remote; 3 = occasional; 4 = probable; 5 = frequent;
and “B1” (potential for new business/products
/services), “B2” (potential for organizational expansion),
“B3” (potential for satisfying regulations), “B4”
(potential for the improvement of QMS processes), and
“B5” (potential improvement of organizational
reputation) were rated as 1 = none/N/A; 2 = minor;
3 = moderate; 4 = high; 5 = very high; while “B6”
(estimated cost of implementation) was rated as
1 = >PhP1M; 2 = PhP500,001–1,000,000;
3 = PhP100,001–500,000; 4 = <PhP100,000; 5 = 0 or
N/A.
The values of “R” and “O”, including their respective
parameters “PR,” “S,” “PO,” and “B,” were evaluated
twice a year. Hence, the notation of subscripts “A” and
“B” was used to indicate the first and second period
ratings, respectively. The following equations were used
to indicate the specific period ratings:
𝑅𝐴 = 𝑃𝑅𝐴𝑆𝐴 (7)
𝑅𝐵 = 𝑃𝑅𝐵𝑆𝐵 (8)
𝑃𝑅𝐴 = [𝑃𝑅1𝐴+𝑃𝑅2𝐴
2] (9)
𝑃𝑅𝐵 = [𝑃𝑅1𝐵+𝑃𝑅2𝐵
2] (10)
𝑆𝐴 = [𝑆1𝐴+𝑆2𝐴+𝑆3𝐴+𝑆4𝐴+𝑆5𝐴+𝑆6𝐴
6] (11)
𝑆𝐵 = [𝑆1𝐵+𝑆2𝐵+𝑆3𝐵+𝑆4𝐵+𝑆5𝐵+𝑆6𝐵
6] (12)
𝑂𝐴 = 𝑃𝑂𝐴𝐵𝐴 (13)
𝑂𝐵 = 𝑃𝑂𝐵𝐵𝐵 (14)
𝑃𝑂𝐴 = [𝑃𝑂1𝐴+𝑃𝑂2𝐴
2] (15)
𝑃𝑂𝐵 = [𝑃𝑂1𝐵+𝑃𝑂2𝐵
2] (16)
𝐵𝐴 = [𝐵1𝐴+𝐵2𝐴+𝐵3𝐴+𝐵4𝐴+𝐵5𝐴+𝐵6𝐴
6] (17)
𝐵𝐵 = [𝐵1𝐵+𝐵2𝐵+𝐵3𝐵+𝐵4𝐵+𝐵5𝐵+𝐵6𝐵
6] (18)
The author designed the FIS module by using
MATLAB. Figure 6 shows the two combined Mamdani-
type FIS-1 models intended to realize objectives 1 and 2
and the FIS-2 model used to achieve objective 3 of the
study. As shown, FIS-1 model is composed of 4 input
data for “R,” namely, “PRA,” “SA,” “PRB,” and “SB”;
another 4 input data for “O,” namely, “POA,” “BA,”
“POB,” and “BB”; 2 output data for “R,” namely, “RA-
RAA,” “RB-RAB”; and 2 output data for “O,” namely,
“OA-OAA” and “OB-OAB.”
To characterize the input and output parameters of the
FIS-1 model, the author designed their respective MF,
as elaborated in Table 2. A sample representation of an
input data “PRA” is further elucidated in Figure 7.
Figure 6. FIS-1 and FIS-2 Models
ISO 9001:2015 Risk-based Thinking using Fuzzy-Support Vector Machine
Makara J. Technol. 1 December 2020 | Vol. 24 | No. 3
153
Table 2. MF for Risks and Opportunities
Data/Role MF Type Parameters Definition/Action Plan
PRA, PRB, POA,
POB
(Input)
Gaussmf [0.25 1] Improbable
[0.25 2] Remote
[0.25 3] Occasional
[0.25 4] Probable
[0.25 5] Frequent
SA, SB, BA, BB
(Input)
Gbellmf [0.25 2.5 1] Negligible/None
[0.25 2.5 2] Minor
[0.25 2.5 3] Serious/Moderate
[0.25 2.5 4] Critical/High
[0.25 2.5 5] Catastrophic/Very High
RA-RAA, RB-RAB
(Output)
Trimf [0 2.5 5] Very Low – Take Risk in Order to Pursue an Opportunity
(VL-TRIOTPAO)
[5.01 7.5 10] Low – Retain Risk by Informed Decision (L-RRBID)
[10.01 12.5 15] Medium – Change Probability or Severity (M-CPOS)
[15.01 17.5 20] High – Eliminate the Risk Source (H-ETRS)
[20.01 22.5 25] Very High – Avoid the Risk (VH-ATR)
OA-OAA, OB-OAB Trimf [0 2.5 5] Very Low – Reject Opportunity Outright (VL-ROO)
[5.01 7.5 10] Low – Consider Opportunity for Further Decisions (L-
COFFD)
[10.01 12.5 15] Medium – Accept Opportunity Under Controlled
Conditions (M-AOUCC)
[15.01 17.5 20] High – Explore in Greater Details Before Pursuing (H-
EGDBP)
[20.01 22.5 25] Very High – Pursue the Opportunity (VH-PTO)
Table 3. MF for Effectiveness of Action Plans
Data/Role MF Type Parameters Action Plan/Effectiveness Level
RA-RAA, RB-RAB
(Input)
Gbellmf [1.25 2.5 2.5] Very Low – Take Risk in Order to Pursue an
Opportunity (VL-TRIOTPAO)
[1.245 2.501 7.5] Low – Retain Risk by Informed Decision
(L-RRBID)
[1.245 2.501 12.5] Medium – Change Probability or Severity
(M-CPOS)
[1.246 2.501 17.5] High – Eliminate the Risk Source (H-ETRS)
[1.245 2.501 22.5] Very High – Avoid the Risk (VH-ATR)
OA-OAA, OB-OAB
(Input)
Gaussmf [1.061 2.5] Very Low – Reject Opportunity Outright
(VL-ROO)
[1.061 7.5] Low – Consider Opportunity for Further Decisions
(L-COFFD)
[1.061 12.5] Medium – Accept Opportunity Under Controlled
Conditions (M-AOUCC)
[1.061 17.5] High – Explore in Greater Details Before Pursuing
(H-EGDBP)
[1.061 22.5] Very High – Pursue the Opportunity
(VH-PTO)
RE, OE Trimf [-46.15 -37.55 -28.94] Very Not Effective (VNE)
[-28.57 -19.29 -10] Not Effective (NE)
[-9.524 0 9.524] Moderately Effective (ME)
[9.999 19.35 28.57] Effective (E)
[28.94 37.55 46.15] Very Effective (VE)
Corpuz
Makara J. Technol. 1 December 2020 | Vol. 24 | No. 3
154
Afterward, the author established the relationship
between the input and output data through 100 sets of
fuzzy rules. The extent of programming the maximum
number of rules was based on the possible logical
combination of inputs vis-à-vis the desired output
response. Each rule has a pair of an antecedent (“if”), a
consequent (“then”), and with a total weight of “1.”
Figure 8 shows a screenshot of sample fuzzy rules (i.e.,
95–100) verbosely coded for risks, opportunities, and
action plans of the FIS-1 model.
Meanwhile, the FIS-2 model was designed with 2 input
data “RA-RAA” and “RB-RAB” for “R” and 2 input data
“OA-OAA” and “OB-OAB” for “O.” It also has 1 output
data “RE” and 1 output data “OE.” Table 3 summarizes
the MF of the input and output data of the FIS-2 model,
while Figure 9 details a sample MF of output data “RE.”
Figure 7. MF-PRA Input Data
Figure 8. FIS-1 Fuzzy Rules
Figure 9. MF- RE Output Data
Figure 10. FIS-2 Fuzzy Rules
Subsequently, the FIS-2 model was programmed with
50 sets of fuzzy rules to accurately determine the level
of effectiveness of “RE” and “OE.” Figure 10 shows a
screenshot of the 50 fuzzy rules (i.e., 45–50) of the FIS-
2 model.
After modeling, the author analyzed 20% of 211
datasets as sample test data and then compared the
performance of the FIS-1 and FIS-2 models with that of
the conventional manual method by using Google
Sheets, where the accuracy and timeliness of both
methods were noted. The comparison results were
analyzed by using paired sample t-test with 95%
confidence level and 5% percentage error on the basis
of the following formulas [10,21]:
H0:μ1=μ2 (19)
𝑡 =�̅�𝑑𝑖𝑓𝑓−0
𝑆�̅� (20)
𝑆�̅� =𝑆𝑑𝑖𝑓𝑓
√𝑛 (21)
where H0 is the null hypothesis; μ1 is the population
mean of the first variable; μ2 is the population mean of
the second variable; t is the test statistic; “x̅diff” is the
sample mean of the differences; n is the sample size;
Sdiff is the sample standard deviation of the differences;
and Sx̅ is the estimated standard error of the mean “S
√n.”
The last module designed was the SVM module, which
was designed by using MATLAB. This module was
used to analyze the text patterns of the previously stated
“Ri” and “Op” in the corpus module. The corresponding
“RE” and “OE” of the FIS-2 model were used as a
reference in defining the feasibility “F,” which was
precomputed by using the following formula and further
interpreted in Table 4.
𝐹 = [𝑅𝐸+𝑂𝐸
2] (22)
In the SVM module, the author loaded and extracted the
input and output data taken from the corpus module,
which were composed of 422 input data “Ri_Op” and
output data “F.” The input data were initially
preprocessed through tokenization or collection of
words for text analysis; filtering of stop words; stemming
Table 4. Feasibility of Risks and Opportunities
Level of Effectiveness Feasibility Interpretation
>+28.94 Very Feasible (VF)
+9.999 to +28.93 Feasible (F)
>−-9.524 to +9.998 Moderately Feasible (MF)
>-28.57 to -9.523 Not Feasible (NF)
>-46.15 to -28.56 Very Not Feasible (VNF)
ISO 9001:2015 Risk-based Thinking using Fuzzy-Support Vector Machine
Makara J. Technol. 1 December 2020 | Vol. 24 | No. 3
155
Figure 11. Tokenized Documents
Figure 12. Bag-Of-Words Model Results
Figure 13. ECOC SVM Classification Model Algorithm
or lemmatization to normalize the texts; removal of
punctuations, short words (<2 characters), and long
words (>15 characters); and the bag-of-words model.
Figures. 11 and 12 show the results of the preprocessing
steps, which resulted in 422 tokenized documents and a
total of 994 vocabularies.
Afterward, the author designed a text classification
model by using a compact version of the multi-class
error correcting output codes (ECOC) for SVM binary
learners with a one-versus-one coding design. Figure 13
shows the algorithm used to train the SVM model in a
supervised learning environment. It employed word
frequency counts of the bag-of-words model “Ri_Op”
as predictors and the feasibility levels “F” as the
response.
After the training, the author evaluated the performance
of the model in terms of classification accuracy “acc,”
which is the proportion of the labels that the model
predicted correctly, by using the following formula:
𝑎𝑐𝑐 = (𝑇𝑃 + 𝑇𝑁)
(𝑇𝑃 + 𝐹𝑃 + 𝐹𝑁 + 𝑇𝑁) (23)
where TP is the number of true positives; TN is the
number of true negatives; FP is the number of false
positives; and FN is the number of false negatives. The
“acc” was further validated using a confusion matrix
wherein the classification error, was determined by the
following equation:
𝐶𝐸 = ∑ 𝑤𝑗𝑒𝑗
𝑛𝑗=1
∑ 𝑤𝑗𝑛𝑗=1
(24)
where wj is the weight for the observation j, which is
normalized by the software to sum to 1; and ej = 1 if the
predicted class of observation j differs from its true
class, and 0 otherwise [22].
The author then simulated the model to test its
classification performance by using 20 textual
statements of “Ri” and “Op”. Similarly, these test data
were preprocessed and then further analyzed by using
paired sample t-test to determine if a significant
difference exists between the target and output classes.
3. Results and Discussion
The proposed three-module framework generally aimed
to document the requirements of the ISO 9001:2015
standard and to improve the efficiency of data collection
and decision-making through AI techniques.
Corpuz
Makara J. Technol. 1 December 2020 | Vol. 24 | No. 3
156
Figure 14. Corpus Module
(a) FIS-2 Input Data
(b) FIS-2 Output Data
Figure 15. FIS-1 Rule Viewer Results
(a) FIS-2 Input Data
(b) FIS-2 Output Data
Figure 16. FIS Rule Viewer Results
The corpus module was designed as a dynamic
collection of textual statements of issues “I,” current
controls “C” to mitigate the issues, and the risks “Ri”
and opportunities “Op.” In this study, the author
analyzed raw and unstructured 211 text-based data sets
as shown on Figure 14. “I” and “C” were used as
references only, while “Ri” and “Op” were practically
used in the succeeding SVM module. These text data
were initially identified by concerned PO. Hence, the
contents could be changed over time.
The FIS module was made up of the FIS-1 model used
to automate the analysis of “Ri” and their action plans
“RA-RAA” and “RB–RAB,” as well as “Op” and their
action plans “OA-OAA” and “OB-OAB”; and the FIS-2
model to evaluate the effectiveness of action plans “RE”
and “OE.” The modeling results are illustrated in the
rule and surface viewers of the MATLAB computing
software. Figure 15 shows a sample rule viewer of the
FIS-1 model, which illustrates how the model generates
the output data depending on the combination of input
data. In this particular example, if the input values are
PRA = 4.05 (probable-to-frequent), SA = 3.35 (serious-
to-critical), PRB = 2.25 (remote-to-occasional),
SB = 2.232 (minor-to-serious), POA = 0 (improbable),
BA = 0 (none), BB = 0 (none) and POB = 0 (none); then
the output values are RA-RAA = 12.8 (medium), RB-
RAB = 2.75 (very low), OA-OAA = 2.5 (very low), and
OB-OAB = 2.5 (very low).
Figure 16 shows a sample rule viewer of the FIS-2
model. In this particular example, if the input values are
RA-RAA = 12.8 (medium), RB-RAB = 2.75 (very low),
OA-OAA = 2.5 (very low), and OB-OAB = 2.5 (very
low), then the output data are RE = 37.4 (very effective;
VE) and OE = −6.57 (moderately effective).
The following surface viewers, which are 3D-based
graphical user interfaces, map out the relationship
between 2 input and 1 output data one calculation at a
time. Figure 17-a shows the surface viewer for the FIS-1
model, which shows inputs PRA = 4.05 (probable-to-
frequent) and SA = 3.35 (serious-to-critical) and the
output RA-RAA = 12.8 (medium). Meanwhile, Figure
17-b shows the surface viewer for the FIS-2 model with
inputs RA-RAA = 12.8 (medium) and RB-RAB = 2.75
(very low), and the output RE = 37.4 (VE). These maps
indicate the extent of the direct or inverse relationship
between 2 inputs and 1 output variable in a particular
period of analysis.
ISO 9001:2015 Risk-based Thinking using Fuzzy-Support Vector Machine
Makara J. Technol. 1 December 2020 | Vol. 24 | No. 3
157
After modeling, the FIS module was tested to evaluate
its accuracy vis-à-vis the performance of the
conventional method. Table 5 shows the results of
paired samples t-test for the FIS-1 model. It reveals that
no significant difference exists in the accuracy between
the two methods, with t and p values of t1(42) = 1.355,
t2(42) = .443, t3(42) = 1.431, and t4(42) = 1.775 with
p1 = .183, p2 = .660, p3 = .160, and p4 = .083.
Interestingly, as elaborated in Table 6, the study found a
statistically significant improvement in the timeliness of
the computation in all test cases from 5.49 s ± 1.099 s to
2.23 s ± .782s (p = .000) for RAA, from 5.47 s ± 1.241 s
to 2.14 s ± .804 s (p = .000) for RAB, from 5.40 s ± 1.11
6 s to 2.21 s ± .742 s (p = .000) for OAA, and from 5.33
s ± 1.169 s to 2.09 s ± .895 s (p = .000) for OAB.
(a) FIS-1 (b) FIS-2
Figure 17. Sample FIS Surface Viewer Results
Subsequently, the FIS-2 model was tested to establish
its accuracy and timeliness in determining the “RE” and
“OE” in comparison with the conventional method. The
results of paired sample t-test, as shown in Table 7,
indicate that no significant difference exists in the
accuracy between the two methods, with t-values of
t1(42) = .443 and t2(42) = −.530, p1 = .660, and p2 =
.599.
Furthermore, Table 8 shows the inferential results of
paired sample t-test on the timeliness performance of
the FIS-2 model. The results reveal a statistically
significant improvement in the timeliness of
computation in all test cases from 2.65 s ± .482 s to 1.44
s ± .782 s (p = .000) for RE, and from 2.51 s ± .506 s to
1.42 s ± .499 s (p = .000) for OE.
Lastly, the SVM module was designed as a decision
support system that would be capable of analyzing the
text patterns of previously identified “Ri” and “Op” and
predicting the feasibility “F” of future proposals based
on related RBT parameters. The training results show
that the SVM model has a 97.16% classification
accuracy, as coded in Figure 18, and a confusion error
rating of 2.6%, as shown in the confusion matrix in
Figure 19.
Table 5. FIS-1 Accuracy Paired Sample T-Test
Mean Std.
Deviation
Std. Error
Mean
95% Confidence Interval
of the Difference t df Sig. (2-tailed)
Lower Upper
Pair 1 RAA
FIS1_RAA .070 .338 .052 -.034 .174 1.355 42 .183
Pair 2 RAB
FIS1_RAB .023 .344 .052 -.083 .129 .443 42 .660
Pair 3
OAA
FIS1_OA
A
.047 .213 .032 -.019 .112 1.431 42 .160
Pair 4 OAB
FIS1_OAB .070 .258 .039 -.010 .149 1.775 42 .083
Table 6. FIS-1 Timeliness Paired Sample T-Test
Mean Std.
Deviation
Std. Error
Mean
95% Confidence Interval
of the Difference t df Sig. (2-tailed)
Lower Upper
Pair 1 RAA
FIS1_RAA 3.256 1.449 .221 2.810 3.702 14.734 42 .000
Pair 2 RAB
FIS1_RAB 3.326 1.443 .220 2.882 3.770 15.114 42 .000
Pair 3 OAA
FIS1_OAA 3.186 1.258 .192 2.799 3.573 16.602 42 .000
Corpuz
Makara J. Technol. 1 December 2020 | Vol. 24 | No. 3
158
Table 7. FIS-2 Accuracy Paired Samples T-Test
Mean Std.
Deviation
Std.
Error
Mean
95% Confidence Interval
of the Difference t df Sig. (2-tailed)
Lower Upper
Pair 1 RE
FIS2_RE .047 .688 .105 -.165 .258 .443 42 .660
Pair 2 OE
FIS2_OE -.093 1.151 .176 -.447 .261 -.530 42 .599
Table 8. FIS-2 Timeliness Paired Sample T-Test
Mean Std.
Deviation
Std. Error
Mean
95% Confidence Interval
of the Difference t df Sig. (2-tailed)
Lower Upper
Pair 1 RE
FIS2_RE 1.209 .742 .113 .981 1.438 10.689 42 .000
Pair 2 OE
FIS2_OE 1.093 .684 .104 .883 1.303 10.485 42 .000
Figure 18. Training Classification Accuracy Results
Figure 19. Training Confusion Matrix Results
After being trained, the SVM model was implemented
and tested using actual 20 datasets. Figure 20 shows a
sample algorithm used to test the SVM model, while
Figure 21 shows the resulting output. The results
indicate that the SVM model was able to predict 19 out
of 20 actual datasets, thereby achieving a 95%
classification accuracy rating during testing.
Figure 20. Testing Algorithm
Figure 21. Results of Testing Algorithm
4. Conclusion
After simulation and statistical testing, the proposed
framework was found to have no significant difference
in terms of accuracy as compared with the conventional
method. However, both FIS-1 and FIS-2 models are
statistically significantly faster than the conventional
method by an average of 3.26 and 1.15 s, respectively.
Moreover, the SVM module has 97.16% classification
accuracy and 2.6% confusion error rating during
training, and an actual classification accuracy of 95%
FM
F NF V
FVNF
Target Class
F
MF
NF
VF
VNF
Ou
tpu
t C
lass
Confusion Matrix
64
15.2%
0
0.0%
0
0.0%
2
0.5%
0
0.0%
97.0%
3.0%
0
0.0%
206
48.8%
0
0.0%
4
0.9%
0
0.0%
98.1%
1.9%
1
0.2%
1
0.2%
40
9.5%
0
0.0%
0
0.0%
95.2%
4.8%
2
0.5%
0
0.0%
0
0.0%
90
21.3%
0
0.0%
97.8%
2.2%
0
0.0%
1
0.2%
0
0.0%
0
0.0%
11
2.6%
91.7%
8.3%
95.5%
4.5%
99.0%
1.0%
100%
0.0%
93.8%
6.2%
100%
0.0%
97.4%
2.6%
ISO 9001:2015 Risk-based Thinking using Fuzzy-Support Vector Machine
Makara J. Technol. 1 December 2020 | Vol. 24 | No. 3
159
during testing. The strength of the SVM module is its
text classification feature for predicting text patterns of
risks and opportunities and their parameters; this ability
is not evident in existing conventional RBT systems.
Hence, all the results affirm that the use of FIS and
SVM is a feasibly efficient approach in designing an
RBT framework in conformance with the requirements
of ISO 9001:2015. Future research should focus on
increasing the RBT data collected from other similar
organizations to enhance the external validity of the
proposed framework.
Acknowledgements
The author would like to thank the support of the top
management and all POs toward the continuous
improvement of the QMS of TUP and the efforts
exerted by the quality assurance staff during the data
collection.
References
[1] R.S.A. Corpuz, Int. J. Recent Technol. Eng. 8
(2019) 420.
[2] International Organization for Standardization, ISO
9001:2015 Quality management systems
requirements, 2015.
[3] International Organization for Standardization, ISO
9001:2015 Fundamentals and vocabulary, 2015.
[4] A.Y. Ezrahovich, A.V. Vladimirtsev, I.I. Livshitz,
P.A. Lontsikh, V.A. Karaseva, International
Conference Quality Management, Transport and
Information Security, Information Technologies,
St. Petersburg, Russia, 2017, p. 506.
[5] International Organization for Standardization, ISO
31000:2009 Risk management: principles and
guidelines, 2015.
[6] T. Aven, Risk Anal. 32 (2012) 1647.
[7] H. Pačaiová, J. Sinay, A. Nagyová, Meas. 100
(2017) 288.
[8] M. Gallab, H. Bouloiz, Y. L. Alaoui, M. Tkiouat,
Second International Conference on Intelligent
Computing in Data Sciences (ICDS), Fez,
Morocco, 2019, p. 226.
[9] Z. Zadeh, Comput., 1 (1988) 83.
[10] A.B. Villanueva, R.S.A. Corpuz, Int. J. Sci.
Technol. Res. 9 (2020) 2096.
[11] L. Pokoradi, AARMS. 1 (2002) 63.
[12] V. Vapnik, The Nature of Statistical Learning
Theory (2nd ed), Springer, Berlin, 1999.
[13] W. Feng, Y. Xiao, Pacific-Asia Conference on
Circuits, Communication and System, Chengdu,
China, 2009, p. 710.
[14] T. Hastie, R. Tibshirani, J. Friedman. The Elements
of Statistical Learning, second edition. Springer,
New York, 2008.
[15] M.Y. Cheng, A.F.V. Roy, K.L. Chen, Expert Syst.
Appl. 39 (2012) 1737.
[16] C. Burges, Data Min. Knowl. Discov. 2 (1998)
121.
[17] Y. Wang, S. Wang, K.K. Lai, IEEE Trans. Fuzzy
Syst. 13 (2005) 820.
[18] P.Y. Hao, M-S. Lin, L-B. Tsai, 8th International
Conference on Intelligent Systems Design and
Applications, Vellore, India, 2008, p. 83.
[19] Z. Jia, L. Gong, J. Han, International Conference
on Computer Science and Software Engineering,
Wuhan, China, 2008, p. 508.
[20] C.K. Lau, K.K. Lai, Y.P. Lee, J. Du, Fire Saf. J. 28
(2015) 188.
[21] R.S.A. Corpuz, J. C. Orquiza, 10th International
Conference on Humanoid, Nanotechnology,
Information Technology, Communication and
Control, Environment and Management (HNICEM),
Baguio City, Philippines, 2018,
[22] MathWorks Inc., Classification ECOC Loss,
https://www.mathworks.com/help/stats/classificatio
necoc.loss.html, 2020.